Apache TVM v0.27.0.rc0 announces deprecation of tirx.ptx.*

The release notes for Apache TVM v0.27.0.rc0, as published on the project's GitHub, include a significant change involving the CUDA backend. The tirx.ptx.* namespace has been retired and replaced with a table-driven PTX dialect. This update aligns the PTX dialect with newer PTX ISA versions, introducing typed PTX forms and enhancing CUDA compatibility. Additionally, support for newer CUTLASS packages and CUDA 13.4 features has been added, along with synchronization improvements and expanded tensor-map capabilities.

Operators planning to upgrade should carefully assess any dependencies or custom implementations that rely on the tirx.ptx.* namespace. The removal of this namespace could impact workflows or scripts that directly reference it, requiring modifications to align with the new PTX dialect. This change reflects a broader pattern in projects where legacy namespaces are deprecated in favor of more structured and extensible approaches. Testing in environments with CUDA 13.4 or newer is recommended to ensure compatibility with the updated features.